| Message ID | alpine.DEB.2.20.1701311521430.3457@nanos (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tue, Jan 31, 2017 at 3:24 PM, Thomas Gleixner <tglx@linutronix.de> wrote: > The handling of the might_cancel queueing is not properly protected, so > parallel operations on the file descriptor can race with each other and > lead to list corruptions or use after free. > > Protect the context for these operations with a seperate lock. > > The wait queue lock cannot be reused for this because that would create a > lock inversion scenario vs. the cancel lock. Replacing might_cancel with an > atomic (atomic_t or atomic bit) does not help either because it still can > race vs. the actual list operation. > > Reported-by: Dmitry Vyukov <dvyukov@google.com> > Signed-off-by: Thomas Gleixner <tglx@linutronix.de> > --- > fs/timerfd.c | 17 ++++++++++++++--- > 1 file changed, 14 insertions(+), 3 deletions(-) > > --- a/fs/timerfd.c > +++ b/fs/timerfd.c > @@ -40,6 +40,7 @@ struct timerfd_ctx { > short unsigned settime_flags; /* to show in fdinfo */ > struct rcu_head rcu; > struct list_head clist; > + spinlock_t cancel_lock; > bool might_cancel; > }; > > @@ -112,7 +113,7 @@ void timerfd_clock_was_set(void) > rcu_read_unlock(); > } > > -static void timerfd_remove_cancel(struct timerfd_ctx *ctx) > +static void __timerfd_remove_cancel(struct timerfd_ctx *ctx) > { > if (ctx->might_cancel) { > ctx->might_cancel = false; > @@ -122,6 +123,13 @@ static void timerfd_remove_cancel(struct > } > } > > +static void timerfd_remove_cancel(struct timerfd_ctx *ctx) > +{ > + spin_lock(&ctx->cancel_lock); > + __timerfd_remove_cancel(ctx); > + spin_unlock(&ctx->cancel_lock); > +} > + > static bool timerfd_canceled(struct timerfd_ctx *ctx) > { > if (!ctx->might_cancel || ctx->moffs != KTIME_MAX) > @@ -132,6 +140,7 @@ static bool timerfd_canceled(struct time > > static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags) > { > + spin_lock(&ctx->cancel_lock); > if ((ctx->clockid == CLOCK_REALTIME || > ctx->clockid == CLOCK_REALTIME_ALARM) && > (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) { > @@ -141,9 +150,10 @@ static void timerfd_setup_cancel(struct > list_add_rcu(&ctx->clist, &cancel_list); > spin_unlock(&cancel_lock); > } > - } else if (ctx->might_cancel) { > - timerfd_remove_cancel(ctx); > + } else { > + __timerfd_remove_cancel(ctx); > } > + spin_unlock(&ctx->cancel_lock); > } > > static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx) > @@ -400,6 +410,7 @@ SYSCALL_DEFINE2(timerfd_create, int, clo > return -ENOMEM; > > init_waitqueue_head(&ctx->wqh); > + spin_lock_init(&ctx->cancel_lock); > ctx->clockid = clockid; > > if (isalarm(ctx)) Can't we still end up with an inconsistently setup timer? do_timerfd_settime executes timerfd_setup_cancel and timerfd_setup as two separate non-atomic actions. So if there are 2 concurrent timerfd_settime calls, one that needs cancel and another that does not need cancel, can't we end up with inconsistent setup? E.g. setup timer that needs cancel, but it won't be in cancel_list. Or vice versa.
On Wed, 1 Feb 2017, Dmitry Vyukov wrote: > > Can't we still end up with an inconsistently setup timer? > do_timerfd_settime executes timerfd_setup_cancel and timerfd_setup as > two separate non-atomic actions. So if there are 2 concurrent > timerfd_settime calls, one that needs cancel and another that does not > need cancel, can't we end up with inconsistent setup? E.g. setup timer > that needs cancel, but it won't be in cancel_list. Or vice versa. Do we really care? If an application arms the timer with cancel in one thread and the same timer without cancel in another thread, then it's probably completely irrelevant whether the state pair timeout/cancel is correct or not. That's clearly an application bug and I don't want to add more locking just to make something which is broken by definition pseudo 'atomic'. Thanks, tglx
On Thu, Feb 2, 2017 at 7:54 PM, Thomas Gleixner <tglx@linutronix.de> wrote: > On Wed, 1 Feb 2017, Dmitry Vyukov wrote: >> >> Can't we still end up with an inconsistently setup timer? >> do_timerfd_settime executes timerfd_setup_cancel and timerfd_setup as >> two separate non-atomic actions. So if there are 2 concurrent >> timerfd_settime calls, one that needs cancel and another that does not >> need cancel, can't we end up with inconsistent setup? E.g. setup timer >> that needs cancel, but it won't be in cancel_list. Or vice versa. > > Do we really care? If an application arms the timer with cancel in one > thread and the same timer without cancel in another thread, then it's > probably completely irrelevant whether the state pair timeout/cancel is > correct or not. That's clearly an application bug and I don't want to add > more locking just to make something which is broken by definition pseudo > 'atomic'. I agree that the program is bogus, and don't have to ensure any sane behavior for it. But I am concerned about potential kernel corruptions due to this. For example, maybe kernel code will decide to not remove such timer from the cancel list on destruction because based on clockid/flags it should not be in the cancel list, but the timer is actually there and we will end up with a leak or a dangling pointer. I did not check that this actually happens, such inconsistent state just looks like a red flag for me.
Dmitry, On Thu, 2 Feb 2017, Dmitry Vyukov wrote: > On Thu, Feb 2, 2017 at 7:54 PM, Thomas Gleixner <tglx@linutronix.de> wrote: > > On Wed, 1 Feb 2017, Dmitry Vyukov wrote: > >> > >> Can't we still end up with an inconsistently setup timer? > >> do_timerfd_settime executes timerfd_setup_cancel and timerfd_setup as > >> two separate non-atomic actions. So if there are 2 concurrent > >> timerfd_settime calls, one that needs cancel and another that does not > >> need cancel, can't we end up with inconsistent setup? E.g. setup timer > >> that needs cancel, but it won't be in cancel_list. Or vice versa. > > > > Do we really care? If an application arms the timer with cancel in one > > thread and the same timer without cancel in another thread, then it's > > probably completely irrelevant whether the state pair timeout/cancel is > > correct or not. That's clearly an application bug and I don't want to add > > more locking just to make something which is broken by definition pseudo > > 'atomic'. > > I agree that the program is bogus, and don't have to ensure any sane > behavior for it. But I am concerned about potential kernel corruptions > due to this. For example, maybe kernel code will decide to not remove > such timer from the cancel list on destruction because based on > clockid/flags it should not be in the cancel list, but the timer is > actually there and we will end up with a leak or a dangling pointer. I > did not check that this actually happens, such inconsistent state just > looks like a red flag for me. That can't happen. ctx->might_cancel and ctx->clist are always in sync with the new lock and that's the only interesting thing. On destruction we don't look at clockid or such, we only care about might_cancel. What is not guaranteed to be in sync is the timer expiry time and the cancel stuff, if two threads operate on the same timerfd in parallel. That's what I do not care about at all. Thanks, tglx
On Fri, Feb 10, 2017 at 11:13 AM, Thomas Gleixner <tglx@linutronix.de> wrote: > Dmitry, > > On Thu, 2 Feb 2017, Dmitry Vyukov wrote: > >> On Thu, Feb 2, 2017 at 7:54 PM, Thomas Gleixner <tglx@linutronix.de> wrote: >> > On Wed, 1 Feb 2017, Dmitry Vyukov wrote: >> >> >> >> Can't we still end up with an inconsistently setup timer? >> >> do_timerfd_settime executes timerfd_setup_cancel and timerfd_setup as >> >> two separate non-atomic actions. So if there are 2 concurrent >> >> timerfd_settime calls, one that needs cancel and another that does not >> >> need cancel, can't we end up with inconsistent setup? E.g. setup timer >> >> that needs cancel, but it won't be in cancel_list. Or vice versa. >> > >> > Do we really care? If an application arms the timer with cancel in one >> > thread and the same timer without cancel in another thread, then it's >> > probably completely irrelevant whether the state pair timeout/cancel is >> > correct or not. That's clearly an application bug and I don't want to add >> > more locking just to make something which is broken by definition pseudo >> > 'atomic'. >> >> I agree that the program is bogus, and don't have to ensure any sane >> behavior for it. But I am concerned about potential kernel corruptions >> due to this. For example, maybe kernel code will decide to not remove >> such timer from the cancel list on destruction because based on >> clockid/flags it should not be in the cancel list, but the timer is >> actually there and we will end up with a leak or a dangling pointer. I >> did not check that this actually happens, such inconsistent state just >> looks like a red flag for me. > > That can't happen. > > ctx->might_cancel and ctx->clist are always in sync with the new lock and > that's the only interesting thing. On destruction we don't look at clockid > or such, we only care about might_cancel. > > What is not guaranteed to be in sync is the timer expiry time and the > cancel stuff, if two threads operate on the same timerfd in > parallel. That's what I do not care about at all. Ack. Thanks for looking at it bearing with me. Then: Acked-by: Dmitry Vyukov <dvyukov@google.com>
On Fri, 10 Feb 2017, Dmitry Vyukov wrote: > On Fri, Feb 10, 2017 at 11:13 AM, Thomas Gleixner <tglx@linutronix.de> wrote: > > ctx->might_cancel and ctx->clist are always in sync with the new lock and > > that's the only interesting thing. On destruction we don't look at clockid > > or such, we only care about might_cancel. > > > > What is not guaranteed to be in sync is the timer expiry time and the > > cancel stuff, if two threads operate on the same timerfd in > > parallel. That's what I do not care about at all. > > Ack. Thanks for looking at it bearing with me. Then: Thanks for asking the questions. It's always good if we need to think it over again. Thanks, tglx
--- a/fs/timerfd.c +++ b/fs/timerfd.c @@ -40,6 +40,7 @@ struct timerfd_ctx { short unsigned settime_flags; /* to show in fdinfo */ struct rcu_head rcu; struct list_head clist; + spinlock_t cancel_lock; bool might_cancel; }; @@ -112,7 +113,7 @@ void timerfd_clock_was_set(void) rcu_read_unlock(); } -static void timerfd_remove_cancel(struct timerfd_ctx *ctx) +static void __timerfd_remove_cancel(struct timerfd_ctx *ctx) { if (ctx->might_cancel) { ctx->might_cancel = false; @@ -122,6 +123,13 @@ static void timerfd_remove_cancel(struct } } +static void timerfd_remove_cancel(struct timerfd_ctx *ctx) +{ + spin_lock(&ctx->cancel_lock); + __timerfd_remove_cancel(ctx); + spin_unlock(&ctx->cancel_lock); +} + static bool timerfd_canceled(struct timerfd_ctx *ctx) { if (!ctx->might_cancel || ctx->moffs != KTIME_MAX) @@ -132,6 +140,7 @@ static bool timerfd_canceled(struct time static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags) { + spin_lock(&ctx->cancel_lock); if ((ctx->clockid == CLOCK_REALTIME || ctx->clockid == CLOCK_REALTIME_ALARM) && (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) { @@ -141,9 +150,10 @@ static void timerfd_setup_cancel(struct list_add_rcu(&ctx->clist, &cancel_list); spin_unlock(&cancel_lock); } - } else if (ctx->might_cancel) { - timerfd_remove_cancel(ctx); + } else { + __timerfd_remove_cancel(ctx); } + spin_unlock(&ctx->cancel_lock); } static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx) @@ -400,6 +410,7 @@ SYSCALL_DEFINE2(timerfd_create, int, clo return -ENOMEM; init_waitqueue_head(&ctx->wqh); + spin_lock_init(&ctx->cancel_lock); ctx->clockid = clockid; if (isalarm(ctx))
The handling of the might_cancel queueing is not properly protected, so parallel operations on the file descriptor can race with each other and lead to list corruptions or use after free. Protect the context for these operations with a seperate lock. The wait queue lock cannot be reused for this because that would create a lock inversion scenario vs. the cancel lock. Replacing might_cancel with an atomic (atomic_t or atomic bit) does not help either because it still can race vs. the actual list operation. Reported-by: Dmitry Vyukov <dvyukov@google.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> --- fs/timerfd.c | 17 ++++++++++++++--- 1 file changed, 14 insertions(+), 3 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html